John Ball

I'm an inventor, scientist and engineer. My passion is in modeling the brains of animals to make machines more useful - with language. We're getting there!

[John Ball @jbthinking]

The risk of using 'AI' that copies other people's code is on the cost-benefit side. You can imagine stacking AI bloatware inside a product will be harder to administer.
And without the generative AI being priced profitably today, lots of errors being introduced into production code-bases may be prohibitively expensive to fix later without ongoing subsidies.

[Proton @ProtonPrivacy]

Hey everyone, we have heard your concerns about Proton’s AI strategy and culture and after internal discussions, we want to share a few things with you.

First, Proton does not install tracking software on employees' laptops to measure how much AI-generated code they produce. This rumor stemmed from a misunderstanding about how MDM software works, and it's not true.

Second, our use of AI will never be at the expense of security. In the past year, contrary to what has been alleged, additional layers of security review have been added to meet the threat posed by hostile AIs scanning Proton's open-source code for vulnerabilities. Our code is open source, and its security remains open for all to verify.

We want to prove this to you, so in the interest of transparency, we will lift the hood on how we use AI internally at Proton. This includes publicly releasing and open sourcing several of the internal tools that we use, including more advanced versions of Lumo with greater capabilities, which have never been made available outside of the company.

We will also publish how code reviews and security reviews at Proton are done, and detail our AI policies, including where AI is permitted, and where it is not, and our rationale. Instead of just being able to see the final open-source code that we produce to verify security, you will also be able to assess how that code was produced.

Lastly, we are not re-interviewing our staff about AI usage, and there are no job losses due to AI re-interviews.

You can hold us to these commitments. We're aware that we haven't lived up to the high standard you ought to hold us to, and for that we can only apologize, we are sorry.

[John Ball @jbthinking]

@TrueAIHound @anilkseth @dwarkesh_sp There is a big difference between science-fiction AI and today’s generative AI. Science-fiction AI is human like, but generative AI is not. Company valuations are better if science-fiction AI is proposed.

[John Ball @jbthinking]

@Grady_Booch If you had a corporate KPI based on lines of code checked in, you could be fired for being too slow. Which highlights the need for a good KPI’s.

[John Ball @jbthinking]

@GaryMarcus And we’re told that AGI is coming within a couple of months at open AI

[John Ball @jbthinking]

I think the problem is framed in two different ways causing the problem. Is AI great if it solves lots of problems? Sure, it’s a tool.

If a CEO claims AI is magic and will solve everything in (random duration) 6 months, 2 years, but in the future. Westerners don’t believe it because it is unlikely based on failed promises in the past.

By now, leaders told us all cars will be autonomous and radiologists are unemployed. The chance of all medicine being solved in the next few months by a statistical tools is unlikely. Very very very unlikely.

2 interpretations for the same thing.

Perhaps CEOs and their friends should focus on what the tools do now, not what they wish they could do in (insert timeframe)

[John Ball @jbthinking]

Exactly. The engineering of ‘machine learning’ has failed at most real-world examples.

Cognitive science holds the key, as by seeing the vast number of issues to solve that engineering today doesn’t handle with the ML/DL approaches.

Science, not engineering alone, is needed for breakthroughs of this nature.

[John Ball @jbthinking]

@TrueAIHound You’d hope that anything bought by customers would always be safe. Even if large like taxis, if they aren’t safe they should be improved until they are.

[John Ball @jbthinking]

The 1000 brains concept assumes a brain is like a parallel computer. But that’s not how animal brains work as we see in damaged brains. A better model sees brains as hierarchical pattern matchers with senses coming in and muscles controlling motion.

Animals don’t need to learn how to move in the same way they don’t learn how to build their body maps. Humans don’t learn to have 5 fingers and 5 toes on each limb.

Learning has a role, but unlike machine learning bodies can do a lot without any learning. Survival doesn’t care why an animal does what it does.

[John Ball @jbthinking]

@TrueAIHound I hear that SpaceX is valuable as they control all commerce in the universe (big) while others are stuck on earth (small). But, isn’t earth the only source of economic activity? 100% of nothing is still a small number.

[John Ball @jbthinking]

@GaryMarcus I hear that in 6-12 months they (humanoid robots) will be sitting in the driver’s seat to control autonomous cars, but I don’t believe it.

[John Ball @jbthinking]

The legacy AI in development today relies on machine learning and statistical techniques that are inefficient and waste enterprise levels of compute and power.

Sending these devices into space may have merit, but their underlying technology is limited and unlike the power of human brains that are the goal of AI.

The better model exploits better ways to do things better. Like us.

Is AI were really progressing, it wouldn’t be targeting text based systems, but instead robots that learn like us from sci fi and that talk to us to get things done without today’s AI chat screen.

[John Ball @jbthinking]

Right. And by definition today’s statistical AI doesn’t understand. A system built on a statistical representation and that chooses based on probability without recognising the world it is manipulating must rely on humans to judge accuracy. The intelligence remains with people.

Real AI can progress rapidly once we turn from following googles transformer model as some kind of breakthrough and focus on what AI is supposed to do - robotics and language understanding is necessary. Simulation of them isn’t the same as duplication. 😀

[John Ball @jbthinking]

The best approach is to use science to progress AI. The use of LLMs integrated into a symbolic model bakes in the energy requirements and limitations of machine learning as a starting point.

That is engineering leading science, not science determining engineering. Without models from the ground up, such dependencies will hold up progress until science is able to progress without preconceptions.

[John Ball @jbthinking]

My experience is that the effort to check code generated by LLMs isn’t worth the effort. It can’t follow SOLID it seems and generates voluminous nonhuman code.

What’s the point in creating a codebase, as the best application of LLMs to date, if it isn’t maintainable? Humans are good at coding even if slower than a machine, but machine code violates the experience of coding.

When I see an enterprise pop up without human intervention that is better than human companies I will be persuaded to look into possibilities more closely. Until then, everything seems to be a year or more away waiting on a couple of once-in-a-lifetime breakthroughs.

In science it is better to do the breakthroughs first since I expect the next breakthrough to be the elimination of the LLM concept completely.

[John Ball @jbthinking]

The weakness in the argument is that today’s AI doesn’t handle the key skills people have. In robotics the inability for driverless cars to equal humans in all conditions. In language the error rate is unacceptable. AI needs to do things, not just provide a statistical possibility for people to check.

[John Ball @jbthinking]

@GaryMarcus It isn’t what you’d expect from insiders expecting an imminent float of $ trillions.

[John Ball @jbthinking]

@rohanpaul_ai The issue with the reasoning is that today’s AI is just a statistical engine. There is no alignment with the goals of AI as seen by lack of progress in human talents like speech, conversation and robotics sensory perception and movement

[John Ball @jbthinking]

@niccruzpatane @SpaceX Funny because that means spacex may be worth $1 per share of which 99 cents is its statistical AI. That’s not good because of the large loss it promises.

[John Ball @jbthinking]

Sure, so which industries have replaced their services with AI automation? Helpdesks? No. Science? No. I can’t see any industry including software development that is replaced by statistical automation from generative AI.

There are fundamental limitations of the LLM model that stop it from providing valuable services without human involvement.

My biggest concern is when the actual cost plus profit is introduced when someone can write a letter with gen AI for $1000 or do it themselves.

Am I missing something in some industry?